DEV Community

AI-KSK
AI-KSK

Posted on Originally published at huggingface.co

Qwen Image: Eros Beta5 FL2VA Ref2VA — 2 Workflows, Read From the Graphs

Qwen Image: Eros Beta5 FL2VA Ref2VA — 2 Workflows, Read From the Graphs

This release ships 2 workflow files built around Qwen Image. Together they total 98 nodes. Rather than summarise the video, this write-up reads the graphs directly — what each one loads, how big its latents are, and which sampler it runs.

The workflows at a glance

Workflow Nodes Links Sampler Latent
10Eros+Beta5+++FL2VA_Ref2VA小脸去油二采精修流+++ 66 114 — —
10Eros+Beta5+SelfLift音频数字人精修流 32 41 euler —

The node count range is wide: 10Eros+Beta5+SelfLift音频数字人精修流 at 32 nodes against 10Eros+Beta5+++FL2VA_Ref2VA小脸去油二采精修流+++ at 66. That gap is usually the difference between a single-pass preview graph and one that adds conditioning, upscale stages, or a second sampling pass.

What the graphs load

Model file Loader node
minimax_h3_latent_upscaler_3d_fp16.safetensors MiniMaxH3LearnedLatentUpscaleT8Advanced
10Eros_Max_h3_TURBO-hybrid_beta5_int8_SinoX.safetensors UNETLoader
minimax_h3_video_vae_fp16.safetensors VAELoader
minimax_h3_audio_vae_fp32.safetensors VAELoader
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors CLIPLoader
qwen3vl_32b_minimax_h3_int8_convrot.safetensors CLIPLoader

LoRAs in play

LoRA Weight
Motion_Repair.safetensors 0.20000000000000004
minimax_h3_lms_v1.0_r64.safetensors 0.30000000000000004
minimax_h3_turbo_4step_10ErosMax_test4_pruned_curveproj1025_exp_v001-T8.safetensors 0.5
MMH3-ExposureSlider-V1.safetensors 0.8
ig_tiktok_aesthetic_h3_lora_v1_500.safetensors 0.5
H3_Combat_V2.safetensors 0.7000000000000002
FrameRush-Minimax-V2_c2-st1500.safetensors 0.5

Not every LoRA runs at 1.0 — the non-default weights here are [0.2, 0.3, 0.5, 0.7, 0.8]. A weight above 1.0 amplifies the LoRA's effect past what its author trained for; below 1.0 backs it off. Both are deliberate tuning choices, not leftovers.

Sampler choice

Samplers used: euler.

A plain euler is deterministic — good when the constraint is the point and noise per step would only perturb the trajectory. Ancestral variants inject noise each step and are the right pick when you want the model to explore rather than commit.

Where this set is weak

  1. Node count is not quality. The largest graph is not automatically the best one; a 40-node graph with a good prompt beats a 200-node graph with a bad one.
  2. No measured comparison in the files. The workflows carry settings, not benchmarks. Any quality claim needs your own A/B run.
  3. Model files are not included. You supply the weights; the graphs only reference filenames.
  4. Version drift. A graph pinned to a specific checkpoint revision will run on a newer one but may produce different output.

Practical order

Open the smallest graph first. It has the fewest moving parts and will confirm your models load before you spend time on the elaborate ones.

Workflows

2 graphs published on RunningHub with invite code rh-v1111:

  • 10Eros+Beta5+++FL2VA_Ref2VA小脸去油二采精修流+++
  • 10Eros+Beta5+SelfLift音频数字人精修流

Files

https://huggingface.co/AI-KSK/ksk-video-307-comfyui-workflows

Video walkthrough: https://youtu.be/KjHrHot5un8
Bilibili: https://www.bilibili.com/video/BV1hEhh6JEvE/

Notes

  • Workflow files only. Model weights are not included and each keeps its own licence.
  • Node counts, geometries and loader filenames above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with the model authors or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (0)